One conventional technique for performing fast keyword/phrase spotting on an audio corpus is to generate a lattice with the lattice entries consisting of phoneme subsequences and store lattice scores and time offsets in a database. At search time, a query is broken down according to which subsequences are included, and all entries for the relevant subsequences are pulled from the database.